Professional Documents
Culture Documents
2 • APRIL 2015
ABSTRACT
We introduce a novel approach to estimating latent oil risk factors and establish
their significance in pricing nonoil securities. Our model, which features four factors
with simple economic interpretations, is estimated using both derivative prices and
oil-related equity returns. The fit is excellent in and out of sample. The extracted oil
factors carry significant risk premia, and are significantly related to macroeconomic
variables as well as portfolio returns sorted on characteristics and industry. The
average nonoil portfolio exhibits a sensitivity to the oil factors amounting to a sixth
(in magnitude) of that of the oil industry itself.
FEW, IF ANY, COMMODITIES have been the focus of more attention for their per-
ceived economic significance than oil. While there is strong evidence relating
oil prices to the business cycle, the nature of the relationship is nonlinear, time-
varying, and difficult to attribute to any single source such as political uncer-
tainty, cartel decisions, or global economic conditions (see Hamilton (2003) and
Barsky and Kilian (2004)). Despite its prominence in the business media and
economics literature, and despite the well-documented role of business cycles in
asset pricing, academic research has largely failed to find consistent evidence
that oil is an important determinant of cross-sectional asset prices.1 This paper
introduces a new model and method to estimate latent oil risk factors using
∗ Ethan Chiang and Keener Hughen are from the Belk College of Business, University of North
Carolina at Charlotte. Jacob Sagi is from the Kenan-Flagler Business School, University of North
Carolina at Chapel Hill. We are grateful to seminar participants at HEC Paris, the University of
Colorado, the University of Florida, the University of Maryland, and the 25th Annual Conference
of the Financial Markets Research Center at Vanderbilt University for their useful comments.
We particularly benefited from comments by Gurdip Bakshi, Bernard Dumas, Pete Kyle, Chris
Leach, Andy Naranjo, Sugata Ray, Georgios Skoulakis, Zijun Wang, the editor (Cam Harvey) and
anonymous referees. Jun Chen provided excellent research assistance. Ethan Chiang and Keener
Hughen acknowledge the support of the Belk College Summer Research Grant. Jacob Sagi thanks
The Nickell Faculty Fund for financial support.
1 Attempts to use Fama and MacBeth (1973) type regressions to deduce oil risk premia from the
cross-section of asset prices have yielded insignificant results (e.g., Chen, Roll, and Ross (1986)).
Direct estimates of risk premia for oil factors using oil derivative price dynamics have yielded mixed
results, but even when they have come up significantly there has been no attempt to connect such
results to cross-sectional asset pricing. Confounding this, Huang, Masulis, and Stoll (1996) find
virtually no relation between futures returns and U.S. stock market returns. On the other hand,
Ferson and Harvey (1993) find that a five-factor model including oil risk performs better than a
single-factor model in predicting the returns of global equity markets. Complementing this finding,
DOI: 10.1111/jofi.12222
769
770 The Journal of FinanceR
noisy price information from both equity and derivative markets. Put simply,
we extract four distinct oil price components and find that these are important
in explaining the movement of asset prices and macroeconomic fundamentals.
While similar models and techniques have been used to explore the valuation
of commodity derivatives, we are the first to tie the methodology to the pric-
ing of equities.2 In particular, the studies that find significant risk premia in
oil-related factors (e.g., Casassus and Collin-Dufresne (2005) and Trolle and
Schwartz (2009)) have not established whether these risk premia are economi-
cally relevant outside oil derivative markets. The oil factors we extract appear
to have an economically and statistically significant relationship with macroe-
conomic variables such as real GDP, industrial production, unemployment,
inflation, and market uncertainty (the VIX). Our four oil factors explain 23.6%
of the variance of oil industry returns beyond what can be explained using mar-
ket returns alone, and this oil risk exposure is associated with a risk premium
of 5.3%. The magnitude of oil risk exposure exhibited by nonoil characteristic
and industry portfolios is on average about a sixth of that of the oil industry,
and compensated by a commensurate risk premium. Overall, our work can be
seen as a new approach to Ross’s (1976) Arbitrage Pricing Theory (APT) em-
ploying structural models together with data from several linked markets to
extract a less noisy set of fundamental and systematic pricing factors.
In our econometric specification, the log-price of oil is affected by four distinct
components. One can be thought of as transient while another may be viewed as
persistent. The drift in the persistent component is itself a factor, reflecting the
long-run trend in oil prices. Finally, the fourth component drives the stochastic
volatility of oil prices (known to exhibit heteroskedasticity). Thus, each of our
model’s components has a ready interpretation in terms of fundamental news
affecting oil prices, which is potentially important in sorting out how oil shocks
impact the macroeconomy (see Hamilton (2003) and Barsky and Kilian (2004)).
We estimate the model using oil futures prices and option-implied variance,
augmented with returns on oil-related stocks. The latter can be an important
source of information about oil prices that may not be captured by short-dated
derivatives. The model is estimated using 30 years of daily data on futures,
options, and equity returns, fitting derivative prices and returns very well
both in and out of sample. The latent factors identified by our estimation
methodology carry significant risk premia, both statistically and economically,
with magnitudes of unconditional annual Sharpe ratios as high as 0.43. Of the
Backus and Crucini (2000) observe that oil shocks are important determinants of relative price
movements in international markets.
2 Related derivative pricing models include Casassus and Collin-Dufresne (2005), Gibson and
Schwartz (1990), Hughen (2010), Schwartz (1997), Schwartz and Smith (2000), and Trolle and
Schwartz (2009). Recent studies linking oil prices and price volatility to oil equities include
Elyasiani, Mansur, and Odusami (2011), Oberndorfer (2009), and Ramos and Veiga (2011). Our
more structural approach to extracting the factors allows us to obtain a better fit to oil equity
prices over a longer horizon, simultaneously fit to oil derivative prices, uncover the risk premia
associated with the oil factors, link oil factors to the macroeconomy, and demonstrate the factors’
relationship with the broader cross-section of equities.
Estimating Oil Risk Factors from Equity and Derivatives Markets 771
four, the oil volatility factor appears to be the most related to nonoil variables,
including the macroeconomic variables mentioned earlier as well as the Fama-
French (2003) factors.
The paper makes several contributions to the literature. First, we demon-
strate that stock prices contain important information about fundamentals,
such as commodity prices, that is not contained in commodity derivatives.
By combining information from both markets, we explain oil-related security
prices and returns better than traditional asset pricing approaches. Second,
our model and the estimation methodology are important to the real options
literature, where valuation relies disproportionately on an accurate descrip-
tion of persistent dynamics and stochastic volatility. Third, we demonstrate
that oil shocks are systematic and command nontrivial prices of risk. Finally,
among the types of news about oil, our approach identifies shocks affecting the
magnitude of uncertainty in oil prices as most important to the macroeconomy
and cross-section of expected returns.
We do not presume that oil (combined with the market) spans all pricing
factors. The paper seeks to provide insights to anyone who believes that oil
plays an important role in the pricing of securities—regardless of whether one
is motivated by ad hoc or general equilibrium considerations to include oil in an
empirical asset pricing model. Our study suggests that stochastic oil volatility
is as important, if not more so, for the macroeconomy than oil price itself.
The study also suggests that permanent and temporary oil price shocks have
different effects on asset prices. In that sense, the paper is best viewed as a
thorough attempt to study how oil risks are linked to oil and nonoil securities
rather than as a prespecified factor model like that of Chen, Roll, and Ross
(1986). Our approach and findings complement recent work that explicitly
models the role of oil in the pricing of securities (see Baker and Routledge
(2012) and Ready (2012, 2013)).
The paper is structured as follows. Section I introduces our four-factor model,
describing its dynamics, derivative prices, and the link between equity returns
and the model risk factors. Section II describes the data we employ and the esti-
mation methodology (Markov chain Monte Carlo, MCMC). Section III presents
the estimation results and the cross-sectional analysis of the extracted latent
factors. Section IV concludes. Appendices A to D contain details of the model
derivation and estimation procedure. A supplementary Internet Appendix con-
tains further information.3
I. Model
Our goal is to estimate oil-relevant state variables from observed prices of
oil futures, options, and equities. To do so efficiently, we look for a specification
admitting closed-form solutions for futures and option prices in terms of fun-
damental state variables. Following the literature on term structure modeling
for interest rates and commodities, we opt for an affine specification (see, e.g.,
3 The Internet Appendix may be found in the online version of this article.
772 The Journal of FinanceR
t = A − A∗ + (B − B∗ )Xt . (1)
A. Futures
The futures price Ft,τ at time t for delivery at time t + τ of one unit of oil is
derived in Appendix B and given by
4 Similar assumptions about the relation between the risk-adjusted and true drifts are made in
Cheridito, Filipovic, and Kimmel (2007), Collin-Dufresne, Goldstein, and Jones (2009), and Hughen
(2010). A dynamic risk premium for oil is consistent with general equilibrium.
5 In Baker and Routledge (2012), the dynamics are a function of risk-sharing among agents
while in Ready (2012) they result from supply fluctuations. Our decomposition better resembles
the supply interpretation in Ready (2012), where long-run, short-run, and volatility fluctuations
in supply can be priced differently in equilibrium. Ready (2013) provides evidence complementary
to ours for the importance of oil supply shocks in the cross-section of returns.
Estimating Oil Risk Factors from Equity and Derivatives Markets 773
δt = r f ,t + κs Xs,t − Xl,t ,
B. Options
The Chicago Mercantile Exchange (CME) lists options on crude oil futures
contracts from which one can hope to extract information about the unspanned
volatility state variable, Xv,t . Calculating model option prices and fitting them
to observed prices using MCMC would be too computationally intensive.
774 The Journal of FinanceR
t Fs,T −s
∞
Ct (T − t, K)e yt,T −t (T −t) − (Ft,T −t − K)+
=2 dK, (3)
0 K2
where Ct (τ, K) is the date t price of a European call option with strike price
K and maturity τ and yt,τ is the time t yield of a risk-free discount bond with
maturity τ . The integral above can be constructed directly from option prices,
and thus the implied variance, IVt,T , is essentially observed (albeit, with error).
In Appendix C, we show that the affine structure of our model leads to a linear
expression of implied variance in Xv,t :
Thus, the information content about Xv,t in option prices can be extracted very
efficiently.
C. Equities
The value of a firm is the present value of all short- and long-term projects.
Thus, the stock prices of firms with high exposure to the oil risk factors should
reflect information about both the long- and short-horizon state variables. Un-
like the case of futures and options, we do not have a structural model of how
stock prices depend on oil prices. Instead, we rely on a simple model that is
consistent with Merton’s (1973) intertemporal CAPM, the general equilibrium
model of Cox, Ingersoll, and Ross (1985), and other continuous-time versions
of Ross’s (1976) APT, where instantaneous stock returns are linear in the un-
derlying state variables. To this end, recall that t from equation (1) is the
instantaneous risk premium of Xt . The absence of arbitrage requires that a
portfolio whose instantaneous returns are perfectly correlated with Xi,t must
have instantaneous expected excess returns of i,t . If our model were perfectly
specified and we observed futures and options prices without error, then we
6 If the horizon T − t is short, then nonzero correlations of futures prices with interest rates will
have a negligible impact. This is the case with the horizon of implied variance data we consider.
Estimating Oil Risk Factors from Equity and Derivatives Markets 775
D. Discretized Model
Data are observed only at discrete times and the transition density of the
state vector must be computed over the length of time between observations.
Because of the stochastic volatility, the transition density for our model is not
known in closed form; however, it can always be approximated by a Gaussian
density by applying an Euler discretization. Although this approximation does
introduce some error in the computed transition density, previous studies show
that the error is typically very small for observation frequencies less than
roughly one week. We use daily observations in our empirical investigation
and thus the error should be small. We also assume that the observed futures
and option prices are measured with white noise error, ξt and et , respectively.
The complete discretized model that we estimate, written in state-space form,
is
√
Xt+t = At + (I + Bt)Xt + tt εt+t , (7)
log Ft,τ = χ (τ ) + ηi (τ )Xi,t + ξt,τ , (8)
i=s,l, p
7 The return on aggregate wealth, for which the market is a proxy, appears in most equilibrium
asset pricing models. Similarly, Chen, Roll, and Ross (1986), Al-Mudhaf and Goodwin (1993),
Sadorsky (2001), and Scholtens and Wang (2008), among others, also include a market factor,
along with proxies for oil risk factors, in their asset pricing models.
776 The Journal of FinanceR
e
rz,t+t = bz,M ReM,t+t + e
bz,i Ri,t+t +
z,t+t , (10)
i=s,l, p,v
where εt , ξt , et , and
t are normal i.i.d.; the coefficient matrices are functions
e
of the parameters and contract maturities; rz,t+t and ReM,t+t denote excess
returns on the equity portfolios indexed by z and on the market, respectively;
e
and Ri,t+t is the i th component of Xt+t − Xt − A∗ t − B∗ tXt , which is the
discrete-time version of the excess return on the factors in equation (5).
8 The actual time to maturity for an oil futures contract is not precisely determined, because
delivery may occur any time during the delivery month at the short side’s discretion. Following
common practice, we assume that delivery occurs at the end of the delivery month. This is because
oil is typically in backwardation, implying that the benefit of ownership exceeds the risk-free rate,
making it advantageous to hold onto the asset for as long as possible. Thus, we assume that the
maturity date is the final business day of the delivery month.
Estimating Oil Risk Factors from Equity and Derivatives Markets 777
average maturities and use all of them in the estimation. In experimenting with this approach,
we found that the term structure of oil-implied variance requires more than the single factor Xv,t
employed in our model. Instead of expanding the model and incorporating more degrees of freedom,
we opted to keep the model structure the same and fit to a single implied variance series. The nearby
series we employ has roughly twice the open interest and trading volume as the second-nearby.
778 The Journal of FinanceR
Table I
Goodness of Fit
The mean daily root mean square errors (RMSEs) for log-futures prices are reported in basis
points, the RMSEs for implied variances are reported in percentages, and the R2 ’s for the four
portfolio returns are reported in percentages. The column labeled “Derivatives Only” reports the
results using the procedure in which the model parameters and latent variables are estimated
using futures and implied variance data and then the factor loadings are estimated from equity
returns. The column labeled “Full Information” reports the results estimated using futures, implied
variances, and returns simultaneously. For comparison, the resulting R2 ’s for the Fama-French-
Carhart (FFC) model are also reported.
A. Estimation
A.1. Goodness of Fit
Table I reports the fit to the data based on the parameter estimates and
filtered state variables. We refer to an estimate of the parameters and latent
variables as “derivatives-only” when the data consist only of the futures and
implied variance series. We refer to the estimate as “full-information” (FI) when
equity returns are also included in the estimation. The parameter estimates
are reported in the Internet Appendix. For the futures, the pricing error is
given by the difference between the fitted log-prices and actual log-prices. Fol-
lowing Trolle and Schwartz (2009), on each day we calculate the root mean
squared error (RMSE) for the futures contracts in the estimation, which is the
square root of the average squared errors across all futures contracts on that
day. The table reports the average of these daily RMSEs. We also calculate an
unadjusted R2 for all 67,545 futures price predictions. Both derivatives-only
and FI estimates deliver excellent fits to the futures data. In both cases the
RMSE is roughly 0.28% and the R2 is essentially one. For comparison, Cortazar
Estimating Oil Risk Factors from Equity and Derivatives Markets 779
and Naranjo (2006) report an RMSE of 0.51% for their Gaussian three-factor
model, Hughen (2010) reports an RMSE of 0.44% for his maximal A1 (3) model,
and Trolle and Schwartz (2009) report an RMSE of 0.39% for the most general
specification of their model; although these studies use different futures data
than we do, the results suggest that misspecification error is not a major con-
cern in our formulation. While the fit is excellent, the RMSE is still about 10
times higher than recent bid-ask spreads on the most liquid contracts.
We also calculate the R2 and RMSE for the implied variance. The RMSE is
calculated as the square root of the time series average of squared errors and
as such understates the goodness of fit relative to the method used by Trolle
and Schwartz (2009).10 While the fit is good (with an R2 of more than 98%
with or without the equity data), it is not as precise as the fit to the futures.
On the other hand, the RMSE is roughly of the order of magnitude of the bid-
ask spread on the associated portfolio of options.11 We conclude that, given
the limitations of the data and the model, the fit to both derivative prices is
excellent.
The most important difference between estimating the model with versus
without the equity data is seen in the fit to equity returns. The proportion
of variance explained by the model for each of the four equity return series
is documented in Table I for both estimation approaches as well as the FFC
model. Using only derivative prices to estimate the parameters yields results
that are on par with the standard FFC model, with an R2 of about 50%. Indeed,
much of the explanatory power in both instances comes from the market compo-
nent, ReM,t , which on average explains 40% of the variance. By contrast, the FI
estimation procedure yields a substantial improvement in explanatory power,
averaging 75.3% of the variation across the four portfolios. This provides strong
evidence that the model is picking up relevant risk factors. The improvement
in R2 is largely driven by improved marginal R2 ’s in the short-term factor Xs,t
and the volatility factor Xv,t (and shared equally between them). We argue later
that this may be because both depend on high-frequency information contained
in equity returns.
Figure 1 plots a time series of the model-estimated factors using FI esti-
mates. The estimation produces a posterior distribution for the value of each
factor on each date and the corresponding figure depicts the mean of that dis-
tribution. Note that the plot for the volatility factor, Xv,t , is hard to distinguish
from observed implied variance over the sample period in which options are
available to construct implied variance. Both the transient and strictly long-
term oil risk components exhibit similar cyclical behavior. This is important
and suggests that the model is capable of disentangling a single shock into dis-
tinct short- and long-term components. The third, persistent, component has
10 To best compare with Trolle and Schwartz (2009), we would have to calculate the average
absolute error for implied variance. Doing so yields a goodness of fit that is comparable to theirs.
11 Using two ticks, or $0.02, for the bid-ask spread on WTI options, a conservative estimate for
∞
the bid-ask spread on the implied variance portfolio is 2 1τ Ft,τ $0.02
K2
dK = 2 $0.02
τ Ft,τ , which averages
over the sample period to roughly 1% (although it would be half as much for recent data).
780 The Journal of FinanceR
Figure 1. Estimated oil factors and option implied variance. This figure depicts time series
of estimated latent oil risk factors at the mean of their posterior distributions: the transient
component, Xs,t (top panel); the long-term component, Xl,t (second panel); the persistent component,
Xp,t (third panel); and unspanned volatility Xv,t (bottom panel, in black and plotted together with
option-implied variance in gray).
Estimating Oil Risk Factors from Equity and Derivatives Markets 781
clearly experienced a secular increase since roughly 2002. This too is intuitive
and predicts that the high price of oil experienced over the past 10 years is
likely to remain high in the foreseeable future. Finally, the volatility compo-
nent appears to experience acute periods of high volatility, reflecting events
causing great uncertainty in energy markets. Most notable among these are
the collapse of the OPEC agreements and subsequent OPEC price wars in late
1985 and early 1986, the First Gulf War between late 1990 and early 1991, the
Second Gulf War in March 2003, and the Global Financial Crisis of late 2008
until early 2009. In addition, Xv,t experienced a high regime between 1999 and
2002 and has since stabilized, albeit at a higher level than in the 1990s. It is
worth noting that the correlation between Xv,t and the GARCH(1,1) volatility
estimate for next-to-nearby futures returns is 0.652 while the correlation be-
tween Xv,t and historical volatility is 0.696. Given that the correlation between
the GARCH(1,1) volatility and historical volatility is a mere 0.463, we conclude
that Xv,t adequately captures stochastic volatility in our estimation approach.
Table II
Subperiod and Out-of-Sample Results
Panel A reports the goodness of fit in subperiods when the model parameters and latent variables
are estimated using futures, option-implied variance, and equities from the full sample. The fit to
the Fama-French-Carhart model is also reported for comparison. Panel B reports in-sample (IS)
and out-of-sample (OOS) goodness of fit in subperiods when the model parameters are estimated
using data on futures, options, and equities from only one of the subperiods (IS). Average daily
root mean square errors (RMSEs) for log-futures prices are reported in basis points, RMSEs for
implied variances are reported in percentages, and R2 ’s for returns are reported in percentages.
Futures RMSE (bps) 27.8 32.0 28.3 30.2 32.6 28.5 29.1 33.1 28.0
ImpVar RMSE (%) 1.5 1.4 1.7 1.7 1.5 1.9 1.6 1.5 1.8
R2 (%)
Small Oil Stocks 33.1 48.9 67.3 27.5 50.8 70.7 30.8 47.2 84.6
Large Oil Stocks 63.5 42.7 83.3 48.9 66.1 86.1 63.9 62.5 87.8
Low B/M Oil Stocks 53.8 56.7 76.3 37.6 71.6 90.5 53.8 66.9 91.4
High B/M Oil Stocks 49.4 53.2 76.2 41.1 60.4 86.6 49.2 57.5 89.6
Average 50.0 50.4 75.8 38.8 62.2 83.5 49.4 58.5 88.4
e
Ri,t+t is the i th component of Xt+t − Xt − A∗ t − B∗ tXt and can therefore be
explicitly calculated from our estimation of the model parameters and latent
e
factors. Panel A of Table III reports time series statistics of the daily Ri,t ’s
estimated with FI. Only the sample mean (i.e., sample risk premium) of Rep,t
is significantly different from zero. This is because, for the size of our sample
period, a series of excess returns must have an annualized in-sample Sharpe
ratio of at least 36% to yield a significant sample mean of excess returns at the
5% confidence level. Only the persistent oil risk factor is associated with an
in-sample Sharpe ratio that is greater than 36%.12 Thus, the relatively short
12 The in-sample Sharpe ratio is obtained by taking the ratio of the daily sample mean and stan-
dard deviation of Rie and multiplying by the square root of 252. To help place this into perspective,
Estimating Oil Risk Factors from Equity and Derivatives Markets 783
Table III
Risk and Return for the Oil Factors and Equity Portfolios
Both panels report statistics based on the full-sample estimates using futures, implied variance,
and equity returns. Panel A reports time series statistics for excess returns on oil factor-mimicking
e
portfolios, Ri,t , defined as the i th component of Xt+t = Xt+t − Xt − A∗ t − B∗ tXt . The first and
third factors correspond to the transient and persistent components in oil prices. The second factor
corresponds to the long-run trend in oil prices, and the last factor corresponds to the stochastic
volatility of oil prices. Panel B reports model estimates of the coefficients from the factor decompo-
sition of returns given by Rte = bM ReM,t + i=s,l, p,v bi Ri,t
e
+
t . The contribution of oil risk factors to
the portfolio’s risk premium is reported in the last column (“Net Oil RP”). Means and confidence
intervals for the net oil risk premia are calculated using the posterior distribution of oil risk premia
estimated from the model. “* ” (“** ”) indicates significance at the 5% (1%) level.
time series restricts one’s ability to directly measure small but economically
meaningful levels of risk premia. Fortunately, we can calculate the risk premia
from the model parameters as follows. Equation (1) identifies the conditional
risk premium in terms of the estimated model parameters and latent factors.
Correspondingly, the unconditional risk premium is given by the four-element
vector.
¯ ≡ A − A∗ + (B − B∗ ) X̄,
where X̄ is the unconditional mean of Xt (under the physical measure) and is a
function of the estimated model parameters only. More importantly, the MCMC
estimation delivers a posterior distribution for ,
¯ thus allowing us to formally
perform significance tests. This is akin to a Fama-MacBeth (1973) regression
where risk premia are also estimated indirectly. In the latter case, sufficient
the FFC portfolios (market excess return, SMB, HML, and UMD) over the same period yielded
annualized in-sample Sharpe ratios of 41%, −2%, 46%, and 54%, respectively.
784 The Journal of FinanceR
13 The negative b ’s are largely robust to changing the priors in the MCMC estimation and to
v
the inclusion of longer-maturity implied variance series (and/or more equity series). Even when
Estimating Oil Risk Factors from Equity and Derivatives Markets 785
options (e.g., Grenadier (2002)), leaving mostly the impact that volatility would
have in the absence of real options.14 In the next subsection we establish that
e
Rv,t is negatively related to indicators of economic growth and positively related
to the VIX index. We speculate that this, together with competition in the oil
industry, is sufficient to overwhelm any real-option effects. In other words, if a
positive shock to Xv,t is sufficiently bad for the entire economy, it is also bad for
the oil industry.
overfitting to the implied variance, boosting the R2 from 98% to essentially one, the bv ’s remain
negative though marginally significant.
14 While there is much concentration in the oil industry, mostly among nationally owned com-
panies in oil producing nations, our focus is on publicly traded international oil companies (IOCs),
which arguably face a more competitive landscape. Since the OPEC price wars of 1986, in addition
to competing with each other, IOCs also face curtailed profits from production and investment
because of governments in consumer countries who appropriate surpluses through taxes (see Bad-
dour (1997)).
15 The VIX data are from the Chicago Board Options Exchange: http://www.cboe.com/
micro/vix/historical.aspx. We use VOX for the pre-1990 period, for which CBOE does not use
the new method to calculate VIX. The other macroeconomic data are from the St. Louis Fed:
http://research.stlouisfed.org/fred2/.
786 The Journal of FinanceR
Table IV
Relationship to Macroeconomic Variables
Panel A reports the correlations, in percentages, between model-extracted oil factor excess returns
and macroeconomic variables. Panel B reports regression coefficients of regressions of macroe-
conomic variables on contemporaneous oil factor excess returns, and the adjusted-R2 ’s of the
regressions. Panel C reports regression coefficients of predictive regressions of macroeconomic
variables on their own lags and lagged oil factor excess returns, and the adjusted-R2 ’s of the
predictive regressions. All regression coefficients are multiplied by 100, and all adjusted-R2 ’s are
reported in percentages. “*” (“**”) indicates significance at the 5% (1%) level. Significance is based
on Newey-West standard errors with 10 lags.
Variable Constant Rse Rle Rep Rve Own Lag Adj R2 (%)
Baa and Aaa rates), change in squared VIX (squares and differences are
e
taken for better comparison with Rv,t ), change in inflation (growth rate of
CPI), real growth of industrial production, change in unemployment, and real
GDP growth. All macroeconomic variable data are annualized quarterly time
Estimating Oil Risk Factors from Equity and Derivatives Markets 787
16 Although some of the macrovariables are available at higher frequencies, we use quarterly
series to better compare the forecasting power of the oil risk factors across the macro variables.
17 We include the lagged macroeconomic variable in the predictive regressions as a regressor to
control for the possibility that xt−1 predicts yt only because yt−1 predicts yt and xt−1 happens to
be correlated with yt−1 . For highly persistent interest rate variables (short rate, term premium,
and default premium), we use first differences (yt ) in the left-hand side, and levels (yt−1 ) in the
right-hand side.
788 The Journal of FinanceR
suggests that oil volatility shocks are not mere proxies for a business cycle
factor.
18 The relationship between Re and HML is unlikely to come from having used B/M-sorted oil
v,t
equity portfolios in our estimation. A portfolio that is long high-B/M oil stocks and short low-B/M
oil stocks has a correlation coefficient of 0.066 with the standard HML factor.
19 Among SMB, HML, and UMD, the average adjusted-R2 (not reported in the table) is 7.3%
with HML having the largest R2 at 14.0% and UMD the smallest at 4.2%.
Table V
Relationship between Fama-French-Carhart and Oil Factors
Panel A reports the coefficients from regressing each oil factor excess return on the Fama-French-Carhart (FFC) factors (SMB, HML, and UMD) and
the market’s excess return (EMKT). Panel B reports the coefficients from regressing each of the FFC factors on the oil factors and market excess
returns. Results of a regression of the market on the oil factor excess returns are also reported. Partial Oil R2 is the difference between the regression
adjusted-R2 and the adjusted-R2 from a regression employing only market excess returns as a regressor. The oil β’s, β j are the regression coefficients
normalized so that the Fama-French oil industry (“FF oil ind”) is defined to have (βs , βl , β p , βv ) = (1, 1, 1, −1). Means and confidence intervals for
the annualized net oil risk premia (reported in the last column) are calculated using the posterior distribution of oil risk premia estimated from the
model. “* ” (“** ”) indicates significance at the 5% (1%) level.
βi (rz ) measures the sensitivity of a portfolio’s returns to oil risk relative to that
of oil industry stocks. Columns 8 to 11 of Panel B, Table V, report the oil β’s
for the SMB, HML, and UMD portfolios. Their magnitude averages to 0.17,
corresponding to about a sixth of the exposure of the oil industry. Placing this
into perspective, the weight of the oil industry portfolio in the market portfolio
is roughly 3.1% throughout our sample period. This supports the notion that
oil risk impacts the economy far beyond oil-related firms.
In the last column of Panel B, Table V, we calculate the unconditional net
oil risk premium for each of the portfolios. This is done using the posterior dis-
tribution of parameters and latent factors produced by the MCMC estimation.
Specifically, we simulate the parameters and Xi,t ’s from the posterior distribu-
tion, regress the FFC and oil industry returns on each corresponding instance
e
of the Ri,t ’s and the market excess return, randomly draw the normally dis-
tributed regression coefficients, and multiply the drawn oil factor regression
coefficients by that simulation’s ¯ to produce a possible realization of the risk
premium that incorporates both model estimation and sample estimation er-
rors. This generates a distribution of risk premia for each portfolio. The table
reports the mean of the distribution and indicates whether it is significantly
different from zero. About a sixth of the risk premia of SMB and HML ap-
pears to be captured by their oil risk exposures. This is consistent with an APT
structure spanned by five to seven fundamental risk factors of which oil is but
one.
Characteristic Portfolios: In Table VI we repeat the exercise of calculating
oil risk β’s (net of the market) and net oil premia for the 25 size-B/M and the
25 size-momentum portfolios from Kenneth French’s data library.20 The oil β’s
documented in Panel B of Table V are reflected in the cross-sectional variation
across the sorted portfolios. Smaller firms (value firms) tend to have a larger
magnitude of exposure than larger firms (growth firms). Recent “winners” have
less negative exposure to oil volatility and more positive exposure to long-term
oil risk than recent losers. All but 3 of the 50 portfolios have negative and
significant exposure to the volatility factor. The average magnitude of exposure
across all 50 portfolios is 20% that of the oil industry, confirming that oil risk
plays a modest but economically significant role in the economy.
As in Table V, the net oil risk premium for each portfolio is calculated at the
bottom (significant premia are in bold to make any pattern easier to identify).
The average magnitude of net oil risk premium across the portfolios is 0.70%
per annum which, while small, is consistent with the hypothesis that oil risk
is one of a handful of key systematic risk factors in the economy. Note that it
is typically larger-firm portfolios that feature significant risk premia despite
20 The two sets of portfolios are constructed differently. The size-B/M portfolios are rebalanced
once per year, while the size-momentum portfolios are constructed daily. This leads to small dif-
ferences in the regression results when aggregating across the B/M and momentum quintiles. In
addition, aggregating over all 25 size-B/M (or size-momentum) portfolios does not result in the
market portfolio that we employ, and when regressed against the market and the oil factors yields
a regression coefficient of 0.89 on the market along with highly significant (albeit small) oil factor
coefficients.
Table VI
β’s and Annualized Percentage Risk Premia for Each Size-Characteristic Sorted Portfolio
The characteristics used are book value (“Book”) and momentum (“Mom”), and the data are taken from Kenneth French’s website. The full-sample
returns on each portfolio are regressed against the excess oil returns and the market to calculate oil β’s, β j . The β j ’s are normalized so that the
Fama-French oil industry is defined to have (βs , βl , β p , βv ) = (1, 1, 1, −1). Means and confidence intervals for the net oil risk premia (reported in the
bottom panel) are calculated using the posterior distribution of oil risk premia estimated from the model. “*” (“**”) indicates significance at the 5%
(1%) level.
Size Size
Book 1 2 3 4 5 Mom 1 2 3 4 5
βs 1 0.25** 0.15** 0.08** −0.01 −0.19** 1 0.35** 0.29** 0.27** 0.23** 0.06*
2 0.25** 0.22** 0.16** 0.16** −0.01 2 0.31** 0.25** 0.20** 0.16** 0.00
3 0.25** 0.27** 0.23** 0.23** 0.12** 3 0.28** 0.24** 0.21** 0.15** 0.00
4 0.27** 0.24** 0.24** 0.24** 0.07** 4 0.28** 0.26** 0.20** 0.15** −0.02
5 0.30** 0.31** 0.21** 0.18** 0.09** 5 0.29** 0.25** 0.22** 0.15** −0.02
βl 1 0.20** 0.10 0.00 −0.10 −0.22** 1 0.30** 0.16 0.02 0.16 −0.06
2 0.24** 0.24** 0.19** 0.19** −0.03 2 0.35** 0.27** 0.21** 0.06 −0.07
3 0.33** 0.36** 0.39** 0.30** 0.17** 3 0.35** 0.37** 0.30** 0.14** 0.00
4 0.36** 0.37** 0.43** 0.31** 0.10 4 0.38** 0.36** 0.31** 0.23** 0.08
5 0.40** 0.50** 0.33** 0.27** 0.29** 5 0.47** 0.42** 0.35** 0.31** 0.20**
βp 1 0.01 −0.05 −0.02 −0.03 −0.12** 1 0.05 0.04 0.07 0.13** 0.10*
2 0.01 0.00 0.01 0.05** −0.01 2 0.06* 0.03 0.06* 0.02 −0.03
3 0.04 0.04 0.09** 0.12** 0.07** 3 0.05* 0.06* 0.07** 0.03* 0.00
4 0.06* 0.06* 0.11** 0.13** 0.07** 4 0.07** 0.03 0.05** 0.05** 0.00
5 0.09** 0.07* 0.07** 0.06* 0.09** 5 0.09** 0.05 0.08** 0.10** 0.05*
βv 1 −0.32** −0.13** −0.05** 0.08** 0.19** 1 −0.49** −0.39** −0.30** −0.24** −0.24**
2 −0.33** −0.25** −0.15** −0.21** −0.04** 2 −0.48** −0.38** −0.33** −0.30** −0.09**
3 −0.35** −0.29** −0.25** −0.32** −0.15** 3 −0.43** −0.35** −0.30** −0.26** −0.11**
4 −0.34** −0.33** −0.34** −0.33** −0.21** 4 −0.39** −0.32** −0.25** −0.19** −0.03**
5 −0.39** −0.36** −0.25** −0.30** −0.14** 5 −0.34** −0.28** −0.18** −0.09** 0.11**
1 0.70 0.10 0.05 −0.24* −0.80* 1 1.16 0.92 0.85 0.93* 0.77**
Estimating Oil Risk Factors from Equity and Derivatives Markets
Net Oil 2 0.71 0.52 0.37 0.59 0.03 2 1.17 0.89 0.85 0.67 0.08
RP (%) 3 0.85 0.76 0.81 1.04* 0.53* 3 1.04 0.89 0.84 0.63 0.21
4 0.89 0.86 1.05 1.09* 0.60* 4 1.01 0.77 0.69 0.57 0.06
791
5 1.10 0.99 0.73 0.81 0.57 5 1.00 0.75 0.66 0.53 −0.06
792 The Journal of FinanceR
having generally lower exposure to oil risk. This is because the larger-firm
portfolios are less volatile and their factor loadings can be determined more
accurately.
The greatest cross-sectional variation in oil factor loadings is between the
size=1 & B/M=5 portfolio and the size=5 & B/M=1 portfolio—among the 25,
it is also these two portfolios that exhibit the highest cross-sectional disper-
sion in expected return. A portfolio long on the former and short on the
latter has an average oil β that is about one-half of that of the oil indus-
try and a net oil risk premium of 1.90%. The size=5 & B/M=1 portfolio
is particularly interesting in that it appears to act as a hedge against oil
risk.
Industry Portfolios: The preceding analysis confirms that our extracted oil
factors are systematic and that portfolios formed on financial characteristics
vary cross-sectionally with respect to oil risk. In particular, firms whose finan-
cial characteristics make them particularly sensitive to business cycle vari-
ables appear to have greater net exposure to oil risk. Next we examine a
different cross-sectional property: industry. The FFC model is not generally
as useful when applied to industry portfolios, possibly because industry port-
folio returns do not vary enough across size and B/M. One does intuitively
expect, however, that some industries will have significantly higher sensitivity
to oil risk than others. To test this intuition, we use the 49 industry returns
collected by Kenneth French and posted on his website, and remove the oil
industry portfolio used earlier in Panel B of Table V. Each day for each indus-
try, starting on June 24, 1983, and ending on December 31, 2012, we regress
the past 60 trading days of industry excess returns on the market excess re-
e
turn and the Ri,t ’s (also from the past 60 trading days). On a daily basis, we
then sort the 48 industry portfolios based on the regression coefficient of Rie
(we do this separately for i = s, l, p, and v). Next, we form eight “octile” port-
folios out of that day’s loading-sorted industry returns (e.g., the lowest octile
equally weights the returns of the six industries with the lowest regression
coefficient). Altogether, we construct 8 × 4 portfolios: eight octiles rebalanced
daily and based on the regression coefficient with respect to each of the four
factors.
If there is no cross-sectional variation in net oil risk exposure across in-
dustries, then each industry would be represented equally in each octile and
appear with frequency 2.08%. Table VII documents the five highest-frequency
industries in the first and last octiles across all days. A clear pattern emerges:
precious metals, mining, and coal consistently appear in octile 8 for Rse , Rle ,
and Rep, and correspondingly in octile 1 for Rve (recall that exposure to Rve is
generally negative for firms that are positively exposed to the other oil fac-
tors). Similarly, construction appears three times and steel twice in these
same octiles. Thus, the industries with the highest oil sensitivity are com-
modities and basic manufacturing. The least oil-sensitive industries appear-
ing consistently (in octile 1 for Rse , Rle , and Rep, and in octile 8 for Rve ) are
tobacco products, candy and soda, computer hardware, defense, and electronic
equipment.
Estimating Oil Risk Factors from Equity and Derivatives Markets 793
Table VII
Industries Most Frequently Represented in the Top and Bottom
Octiles, Sorted Based on Exposure to the Oil Factors
Each day, the excess returns of each of the 48 nonoil industries based on the classification in
Kenneth French’s website are regressed on the four oil factors’ excess returns and the market,
using the previous 60 trading days’ returns. The 48 industries can then be sorted into eight octiles
e
according to the regression coefficient on Ri,t : the six industries with the highest (lowest) recent
e
past sensitivity to Ri,t are equally weighted at date t for the purpose of calculating the date t return
of the eighth (first) octile sorted by sensitivity to Xi .
Xs Xl Xp Xv
Industry Freq (%) Industry Freq (%) Industry Freq (%) Industry Freq (%)
Agric Agriculture
Autos Automobiles and Trucks
Beer Beer & Liquor
Soda Candy & Soda
Coal Coal
Softw Computer Software
Hardw Computers
Cnstr Construction
Guns Defense
Chips Electronic Equipment
FabPr Fabricated Products
Mines Nonmetallic and Industrial Metal Mining
Gold Precious Metals
Ships Shipbuilding, Railroad Equipment
Steel Steel Works Etc
Smoke Tobacco Products
Table VIII
794
β’s and Annualized Percentage Risk Premia for Industries Sorted on Sensitivity to Oil Factors
Each day, the excess returns of each of the 48 nonoil industries based on the classification in Kenneth French’s website are regressed on the four oil
factors’ excess returns and the market, using the previous 60 trading days’ returns. The 48 industries can then be sorted into eight octiles according
e
to the regression coefficient on Ri,t : the six industries in each octile are equally weighted at date t for the purpose of calculating the date t return of
the eighth (first) octile sorted by sensitivity to Xi . This is done daily for each of the Xi ’s to produce 32 octile portfolios. The full-sample returns on
each octile portfolio are then regressed against the excess oil returns and the market to calculate oil β’s, β j . To facilitate comparison, we normalize
the β j ’s so that the Fama-French oil industry is defined to have (βs , βl , β p , βv ) = (1, 1, 1, −1). Means and confidence intervals for the net oil risk
premia (reported in the bottom panel) are calculated using the posterior distribution of oil risk premia estimated from the model. “*” (“**”) indicates
significance at the 5% (1%) level.
Sensitivity to 1 2 3 4 5 6 7 8 8 minus 1
Table VIII reports the net oil β’s and risk premia of the 32 sensitivity-sorted
portfolios. The table confirms that sorting based on the sensitivity of one factor
is highly correlated with sorting based on another factor (negatively correlated,
when sorting on Rve sensitivity). In other words, most firms that are highly
sensitive to one risk factor tend to also be proportionately sensitive to the
others. Moreover, only the top two octiles (bottom two octiles when sorting on
Rve sensitivity) admit a high magnitude of sensitivity to the oil factors. Thus, in
practice, about 25% of industries are substantially affected by oil risk, although
which industries are affected can vary over time. In both Tables VI and VIII,
the preponderance of positive β’s with respect to the first three factors and
negative β’s for the volatility factor suggest that relatively few stocks can be
used to hedge the oil risk factors. Thus, because oil risk is hard to hedge or
diversify away completely, it should be priced. Indeed, the average magnitude
of the β’s is 14.3% while the average magnitude of risk premium is 0.70%.
Averaging the magnitude of exposures of the 32 industry octile portfolios and
the 50 characteristic portfolios from Table VI yields an average magnitude of
exposure that is roughly one-sixth of that of the oil industry. Finally, despite
the fact that only nonoil industries are used to construct the octile portfolios,
a portfolio long octile 8 and short octile 1 (the “8 minus 1” column) will have
oil β’s that are roughly two-thirds the magnitude of the oil industry and a
commensurately high net oil risk premium. Moreover, while each of the “8
minus 1” portfolio risk premia are significant at the 1% level, the hypothesis
that all four have a net oil risk premium of zero can be rejected at the 0.1%
level (accounting for their correlations), consistent with the more demanding
threshold in Harvey, Liu, and Zhu (2013).
In the Internet Appendix we report individual industry and oil subindus-
try loadings and risk premia. Sixteen out of 49 exhibit significant risk
premia with commodity-linked industries having the largest positive risk pre-
mia while retail, technology, and pharmaceuticals have the most negative.
Among oil subindustries, airlines have a large and negative exposure with
a net oil risk premium of −4.88% coming mostly from the persistent fac-
tor. Oil equipment and services is the most sensitive subindustry with 1.80
times the volatility exposure of the oil industry and a net oil risk premium
of 8.24%.
IV. Conclusion
We study the implications of commodity derivative models for equity mar-
kets. We do this by including oil equity returns in the estimation of a structural
model of the term structure of oil futures with unspanned volatility. The esti-
mation technique delivers a time series for four oil risk factors, corresponding
to shocks influencing oil prices in the short-term, in the long-term, persistently,
or through volatility shocks. Our fit to derivative prices is excellent and robust,
and the risk factors we identify appear to be associated with a significant risk
premium. When information from oil-sensitive equities is incorporated into the
estimation of the oil factors, their ability to explain oil-related stock returns
drastically improves (the R2 increases by an average of roughly 20%) while the
fit to derivatives prices remains largely unchanged. Moreover, the additional
information from equities is crucial in allowing us to better estimate the oil
factor risk premia.
The extracted oil risk factors, and in particular the one corresponding to un-
spanned volatility, appear to be systematic. Oil volatility exhibits an econom-
ically and statistically significant negative relationship with growth variables
such as real GDP and industrial production, and a positive relationship with
negative indicators such as the VIX index and unemployment. The magnitude
of oil risk exposure for an average nonoil firm is between 14% and 20% of the
oil risk exposure for the oil industry itself, and corresponds to an average risk
premium magnitude of 0.70%.
Overall, our results provide evidence that oil risk (or alternatively, energy
risk), adequately filtered from noisy price data, is one of a handful of fundamen-
tal factors that affect the pricing of both oil and nonoil securities. In principle,
our methodology could be applied to extracting other fundamental risk factors,
paving the way for a more structured approach to the pricing of securities in
an APT framework.21
Initial submission: May 14, 2012; Final version received: July 10, 2014
Editor: Campbell Harvey
21 Because similar issues have been raised in the literature on currency risk (e.g., Dumas and
Solnik (1995) and Campbell, Serfaty-de Medeiros, and Viceira (2010)), this might be a natural
candidate for an exploration using our method.
Estimating Oil Risk Factors from Equity and Derivatives Markets 797
log St = a + b Yt , (A1)
t t = 0 + 1 × c Yt . (A3)
A quick count reveals there are indeed 28 free parameters (two in A∗ , eight in
B∗ , and nine in each of 0 and 1 since 0 (1, 1) = 0 and 1 (1, 1) = 1), but ad-
missibility will impose restrictions on some of these parameters. In particular,
the drift of the volatility Y4,t must be nonnegative for the volatility to remain
positive, and therefore k41 = k42 = k43 = 0. In addition, when the volatility state
variable Y4,t approaches zero, its instantaneous variance and covariance with
798 The Journal of FinanceR
the other state variables must vanish, otherwise the volatility would drop be-
low zero with positive probability, and therefore the fourth row and column of
0 are zero.
For the purposes of this study, it is more convenient to consider the equivalent
representation
t t = 0 + 1 X4,t ,
and the fourth row and column of 0 are zero. A straightforward calculation
shows that, given such a state variable X, the state variable given by
⎡ ⎤ ⎡ ⎤
0 1 0 1 0
⎢ 0 ⎥ ⎢ −κ1 1 0 0 ⎥
Yt = ⎢ ⎥ ⎢ ⎥
⎣ μ1 ⎦ + ⎣ κ 2 −κ2 κ4 0 ⎦ Xt
1
0 0 0 0 1
1
η̇ = η B∗ + (0, 0, 0, η 1 η) (B2)
2
with initial values χ (0) = 0 and η(0) = (1, 0, 1, 0).
Estimating Oil Risk Factors from Equity and Derivatives Markets 799
The model will exhibit USV if and only if the fourth component η4 is identi-
cally zero, which from equation (B2) is equivalent to the condition
1 1
(η1 , η2 , η3 ) · (κ5 , κ1 κ5 , − − κ5 ) + (η1 , η2 , η3 , 0)1 (η1 , η2 , η3 , 0) ≡ 0. (B3)
2 2
1 − e−κ2 τ
η1 (τ ) = e−κ1 τ , η2 (τ ) = , η3 (τ ) = 1;
κ2
where s11 + 2s13 + s33 = 0. This restriction implies that the characteristic poly-
nomial of 0 has no negative root, that is, 0 is positive semidefinite, if and
only if
⎡ ⎤
s11 s12 −s11 0
⎢ s12 s22 −s12 0⎥
⎢
0 = ⎣ ⎥.
−s11 −s12 s11 0⎦
0 0 0 0
Furthermore, 0 and 1 will have disjoint null spaces if and only if the upper
left 2 × 2 submatrix of 0 is strictly positive definite. See the Internet Appendix
for more details.
22 Many other studies also restrict the risk-neutral “mean reversion” matrix to have a zero
eigenvalue; see Gibson and Schwartz (1990), Schwartz and Smith (2000), and the USV version of
the model in Hughen (2010).
800 The Journal of FinanceR
where −1 < ρ1 , ρ2 < 1, κ1 , κ2 , κ3 , σ1 , σ2 , σ3 > 0, and 2μ2 > σ32 . Trivial relabeling
of the parameters leads to the set of expressions for the model and futures
prices in Section I.
and therefore
T
(T − t) × IVt,T = η(T ∗ − s)0 η(T ∗ − s) ds
t
T
+ η(T ∗ − s)1 η(T ∗ − s) Et∗ [Xv,s ] ds.
t
Because the returns in each period are linear functions of the current and
lagged state variables, it is convenient to consider the stacked state vector
Xt
X̃t = [ ], so that the transition and observation equations become
Xt−t
At I + Bt 0 √ t 0
X̃t+t = + X̃t + t
˜ (D1)
0 I 0 0 0 t+t
and
⎡ ⎤ ⎡ ⎤
χt ηt 0
Zt = ⎣ γt ⎦ + ⎣ δ̃t 0 ⎦ X̃t + ξ̃t , (D2)
bM ReM,t − boil A∗ t
boil
−boil (I + B∗ t)
The first kernel p(IVt |Xv,t , ) is the (Gaussian) likelihood of the implied vari-
ance data viewed as a function of Xv,t , and the second and third kernels
p(rte |Xt , Xt−t , ) and p(rt+te |Xt , Xt+t , ) are the (Gaussian) likelihoods of the
return data on date t and t + t, viewed as functions of Xv,t , with all other
parameters, the values of the other state variables, and the volatility state
variable at all other times treated as fixed. The fourth and fifth kernels
are the transition densities over the current and previous periods, respec-
tively. As a function of Xv,t the conditional density is not a recognizable
distribution; the volatility state variable Xv,t is drawn using a Metropolis-
Hastings independence sampling algorithm with the Gaussian proposal density
e
p(IVt |Xv,t , ) p(rte |Xt , Xt−t , ) p(rt+t |Xt , Xt+t , ) p(Xt |Xt−t , ). If a negative candi-
date for Xv,t is drawn, it is simply discarded and a new candidate is drawn.
The parameter vector is decomposed into five blocks: 1 = (A, B), 2 =
(μl , μv ), 3 = (κs , κl , κv , σs , σl , σv , ρsl , ρvp), 4 = (bM , boil ), and 5 = W is the vari-
ance matrix of the pricing errors. The first block 1 appears only in the drift
term of the state vector and appears linearly; therefore, its conditional density
is Gaussian and may be drawn in one step from a Gaussian distribution using
the Gibbs sampler. Similarly, because the second block 2 appears linearly in
χt , γt , and the risk-neutral drift terms A∗ and B∗ , it appears linearly in the
observation equation equation (D2) and thus it may also be drawn in a single
step directly from a Gaussian distribution. Similarly, the factor loadings bM
and boil appear linearly in equation (D2) and thus the fourth block 4 may also
be drawn in a single step directly from a Gaussian distribution. The fifth block
5 may also be drawn in a single step. We assume homoskedastic variance of
the futures pricing error, the implied variance error, and the variance of the
idiosyncratic return error. By a standard result, the likelihood function of W
is an Inverse Gamma density function and therefore this block may be drawn
directly using Gibbs sampling.
The third block is more difficult to draw because these parameters appear
nonlinearly in equation (D2). As a result, the conditional density of 3 is not
a recognizable density function. This block is drawn using the relatively inef-
ficient random walk Metropolis-Hastings algorithm with a Gaussian proposal
density.
REFERENCES
Al-Mudhaf, Anwar, and Thomas H. Goodwin, 1993, Oil shocks and oil stocks: Evidence from the
1970s, Applied Economics 25, 181–190.
Estimating Oil Risk Factors from Equity and Derivatives Markets 803
Ang, Andrew, Monika Piazzesi, and Min Wei, 2006, What does the yield curve tell us about GDP
growth? Journal of Econometrics 131, 359–403.
Backus, David K., and Mario J. Crucini, 2000, Oil prices and the terms of trade, Journal of
International Economics 50, 185–213.
Baddour, J. W., 1997, The international petroleum industry: Competition, structural change and
allocation of oil surplus, Energy Policy 25, 143–157.
Baker, Steven D., and Bryan R. Routledge, 2012, The price of oil risk, Working paper, Carnegie
Mellon University.
Barsky, Robert, and Lutz Kilian, 2004, Oil and the macroeconomy since the 1970s, Journal of
Economic Perspectives 18, 115–134.
Britten-Jones, Mark, and Anthony Neuberger, 2000, Option prices, implied price processes, and
stochastic volatility, Journal of Finance 55, 839–866.
Campbell, John Y., Karine Serfaty-de Medeiros, and Luis M. Viceira, 2010, Global currency hedg-
ing, Journal of Finance 65, 87-121.
Carhart, Mark M., 1997, On persistence in mutual fund performance, Journal of Finance 52,
57–82.
Casassus, Jaime, and Pierre Collin-Dufresne, 2005, Stochastic convenience yield implied from
commodity futures and interest rates, Journal of Finance 60, 2283–2331.
Chen, Nai-Fu, Richard Roll, and Stephen A. Ross, 1986, Economic forces and the stock market,
Journal of Business 59, 383–403.
Cheridito, Patrick, Damir Filipovic, and Robert L. Kimmel, 2007, Market price of risk specifications
for affine models: Theory and evidence, Journal of Financial Economics 83, 123–170.
Collin-Dufresne, Pierre, Robert S. Goldstein, and Chirstopher S. Jones, 2009, Can interest rate
volatility be extracted from the cross section of bond yields? Journal of Financial Economics
94, 47–66.
Cortazar, Gonzalo, and Lorenzo Naranjo, 2006, An N-factor Gaussian model of oil futures prices,
Journal of Futures Markets 26, 243–268.
Cox, John C., Jonathan E. Ingersoll, and Stephen A. Ross, 1985, An intertemporal general equi-
librium model of asset prices, Econometrica 53, 363–384.
de Jong, Piet, and Neil Shephard, 1995, The simulation smoother for time series models, Biometrika
82, 339–350.
Dumas, Bernard, and Bruno Solnik, 1995, The world price of foreign exchange risk, Journal of
Finance 50, 445–479.
Elyasiani, Elyas, Iqbal Mansur, and Babatunde Odusami, 2011, Oil price shocks and industry
stock returns, Energy Economics 33, 966–974.
Fama, Eugene F., and Kenneth R. French, 1993, Common risk factors in the returns on stocks and
bonds, Journal of Financial Economics 33, 3–56.
Fama, Eugene F., and James D. MacBeth, 1973, Risk, return, and equilibrium: Empirical tests,
Journal of Political Economy 81, 607–636.
Ferson, Wayne E., and Campbell R. Harvey, 1993, The risk and predictability of international
equity returns, Review of Financial Studies 6, 527–566.
Gibson, Rajna, and Eduardo S. Schwartz, 1990, Stochastic convenience yield and the pricing of oil
contingent claims, Journal of Finance 45, 959–976.
Grenadier, Steven R., 2002, Option exercise games: An application to the equilibrium investment
strategies of firms, Review of Financial Studies 15, 691–721.
Hamilton, James D., 2003, What is an oil shock? Journal of Econometrics 113, 363–398.
Harvey, Campbell R., Yan Liu, and Heqing Zhu, 2013, And the cross-section of expected returns,
Working paper, Duke University.
Huang, Roger D., Ronald W. Masulis, and Hans R. Stoll, 1996, Energy shocks and financial markets,
Journal of Futures Markets 16, 1–27.
Hughen, W. Keener, 2010, A maximal affine stochastic volatility model of oil prices, Journal of
Futures Markets 30, 101–133.
Jiang, George J., and Yisong S. Tian, 2005, The model-free implied volatility and its information
content, Review of Financial Studies 18, 1305–1342.
804 The Journal of FinanceR
Liew, Jimmy, and Maria Vassalou, 2000, Can book-to-market, size and momentum be risk factors
that predict economic growth? Journal of Financial Economics 57, 221–245.
Merton, Robert C., 1973, An intertemporal capital asset pricing model, Econometrica 41, 867–887.
Oberndorfer, Ulrich, 2009, Energy prices, volatility, and the stock market: Evidence from the
Eurozone, Energy Policy 37, 5787–5795.
Petkova, Ralitsa, and Lu Zhang, 2005, Is value riskier than growth? Journal of Financial Eco-
nomics 78, 187–202.
Ramos, Sofia B., and Helena Veiga, 2011, Risk factors in oil and gas industry returns: International
evidence, Energy Economics 33, 525–542.
Ready, Robert C., 2012, Oil prices and long-run risk, Working paper, University of Rochester.
Ready, Robert C., 2013, Oil prices and the stock market, Working paper, University of Rochester.
Ross, Stephen A., 1976, The arbitrage theory of capital asset pricing, Journal of Economic Theory
13, 341–360.
Sadorsky, Perry, 2001, Risk factors in stock returns of Canadian oil and gas companies, Energy
Economics 23, 17–28.
Scholtens, Bert, and Lei Wang, 2008, Oil risk in oil stocks, Energy Journal 29, 89–111.
Schwartz, Eduardo, 1997, The stochastic behavior of commodity prices: Implications for valuations
and hedging, Journal of Finance 52, 922–973.
Schwartz, Eduardo, and James E. Smith, 2000, Short-term variations and long-term dynamics in
commodity prices, Management Science 46, 894–911.
Trolle, Anders, and Eduardo Schwartz, 2009, Unspanned stochastic volatility and the pricing of
commodity derivatives, Review of Financial Studies 22, 4423–4461.
Supporting Information
Additional Supporting Information may be found in the online version of this
article at the publisher’s website:
Appendix S1: Internet Appendix